Image recognition simulation model and system
By designing an optical diffraction neural network architecture with a three-layer cascaded metasurface, the computational bottleneck of traditional deep learning models and the large system size of optical diffraction neural networks are solved, enabling high-speed and low-power image recognition task processing and improving feature abstraction capabilities and optical path stability.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- PHOTON ARITHMETIC(BEIJING)TECH CO LTD
- Filing Date
- 2026-03-16
- Publication Date
- 2026-06-05
AI Technical Summary
Traditional deep learning models based on electronic computing suffer from computational bottlenecks and high energy consumption when performing large-scale matrix operations. Furthermore, the multi-layer diffraction unit stacking structure of existing optical diffraction neural networks results in a large system size and poor optical path stability.
An optical diffraction neural network architecture based on a three-layer cascaded metasurface was designed. The nanostructure enables high-degree-of-freedom control of the amplitude, phase, and polarization of the light field in a two-dimensional plane. By extending the network depth through multi-layer cascading, nonlinear mapping is achieved, and the classified light intensity distribution is directly output.
It achieves high-speed, low-power image recognition task processing, improves feature abstraction capabilities, simplifies system structure, and enhances optical path stability and processing efficiency.
Smart Images
Figure CN122156909A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of artificial intelligence technology, and in particular to an image recognition simulation model and system. Background Technology
[0002] With the rapid development of artificial intelligence and image recognition technologies, traditional deep learning models based on electronic computing face challenges in terms of computing power and energy consumption. Electronic computing systems based on the von Neumann architecture require frequent data transfer during large-scale matrix operations, limiting processing speed and energy efficiency, and exhibiting high latency and high power consumption. Optical computing, with its high parallelism and ultra-low latency properties of photons, has become an important direction for overcoming computing power bottlenecks.
[0003] Among them, the optical diffractive deep neural network (D... 2 NN) achieves all-optical inference by directly modulating the phase of the optical field through a multi-layer diffraction structure, enabling tasks such as image classification and target detection without photoelectric conversion, providing a new paradigm for high-speed, low-power intelligent task processing. However, the controllability of the optical field by a single-layer diffraction unit is limited, and existing solutions often employ multi-layer diffraction unit stacking structures, resulting in problems such as large system size and poor optical path stability. Summary of the Invention
[0004] In view of this, the purpose of this invention is to provide an image recognition simulation model and system, which designs an optical diffraction neural network architecture based on a three-layer cascaded metasurface, suitable for high-speed, low-power image recognition.
[0005] In a first aspect, embodiments of the present invention provide an image recognition simulation model, comprising: an optical diffraction neural network for handwritten digit recognition tasks, the optical diffraction neural network comprising: an input layer, a diffraction layer, and an output layer, the input layer, the diffraction layer, and the output layer being connected sequentially; the input layer being used to modulate incident light into a wavefront array carrying image information; the diffraction layer being used to diffract the light output from the input layer multiple times to simulate a convolution weight matrix, adjusting the phase distribution through error backpropagation; and the output layer being used to generate a focused light spot corresponding to the number of categories, the focused light spot focusing on the digits corresponding to multiple detection points, and the intensity distribution of the focused light spot on the detection surface being used to characterize the classification probability.
[0006] In an optional embodiment of this application, the aforementioned optical diffraction neural network is used to project the information carried by the input image onto a preset target detection unit through training, thereby suppressing the light intensity projected onto non-target detection units.
[0007] In an optional embodiment of this application, the metasurface of the optical diffraction neural network includes: a plurality of subwavelength nanostructures, each subwavelength nanostructure serving as a diffraction neuron for training phase values.
[0008] In an optional embodiment of this application, the output of each layer of diffractive neurons is the input of the next layer of diffractive neurons, and the light field and corresponding field distribution of the input metasurface from the optical diffractive neural network are modulated by multiple diffractive neurons.
[0009] In an optional embodiment of this application, the amplitude and phase of the secondary light wave output by each layer of diffractive neurons are determined based on the incident light field of that layer of diffractive neurons and the transmission coefficient of that layer of diffractive neurons.
[0010] In an optional embodiment of this application, the illumination source of the optical diffraction neural network is a complex value, which carries the phase and amplitude information of the input image. The light emitted by the illumination source is diffracted by free space light propagation and serves as the first layer input wave of the optical diffraction neural network.
[0011] In an optional embodiment of this application, the optical diffraction neural network includes: multiple diffraction layers; the input wave of the target diffraction layer is determined based on the light field distribution of the target diffraction layer, the transmission coefficient of the target diffraction layer, and the output wave of the target diffraction layer; the output wave of the target diffraction layer is determined based on the input waves of multiple diffraction layers preceding the target diffraction layer.
[0012] In an optional embodiment of this application, the aforementioned optical diffraction neural network is used to determine the cost function based on the mean square error between the intensity of the focused spot output by the output layer and the target.
[0013] In an optional embodiment of this application, the above-mentioned image recognition simulation model further includes: a sensor array and a fully connected layer; the sensor array is used to receive the optical signal of the focused spot output by the output layer and convert the optical signal into a numerical signal; the fully connected layer is used to determine the probability of the numerical signal belonging to each category through an activation function.
[0014] Secondly, embodiments of the present invention also provide an image recognition simulation system, which includes the image recognition simulation model described above.
[0015] The embodiments of the present invention bring the following beneficial effects: This invention provides an image recognition simulation model and system. The input layer modulates the incident light into a wavefront array carrying image information. The diffraction layer performs multiple diffractions on the light output from the input layer to simulate the convolution weight matrix, adjusting the phase distribution through backpropagation of errors. The output layer generates a focused light spot corresponding to the number of categories. This focused light spot focuses on the numbers corresponding to multiple detection points, and the intensity distribution of the focused light spot on the detection surface is used to characterize the classification probability. In this approach, high-degree-of-freedom control of the light field amplitude, phase, and polarization can be achieved in a two-dimensional plane through nanostructure design. Simultaneously, the network depth is extended through multi-layer cascading to enhance feature abstraction capabilities. Multi-layer cascaded diffraction is used to achieve nonlinear mapping, directly outputting the classification light intensity distribution, which characterizes the classification result.
[0016] Other features and advantages of this disclosure will be set forth in the following description, or some features and advantages may be inferred from the description or determined without doubt, or may be learned by practicing the techniques described above.
[0017] To make the above-mentioned objects, features and advantages of this disclosure more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description
[0018] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.
[0019] Figure 1 This is a schematic diagram of the structure of an optical diffraction neural network provided in an embodiment of the present invention; Figure 2 A schematic diagram illustrating the implementation process of an optical diffraction neural network based on a three-layer cascaded metasurface, provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of an all-optical diffraction neural network provided in an embodiment of the present invention; Figure 4 This is a schematic diagram illustrating the training of an image recognition algorithm based on an optical diffraction neural network with a three-layer cascaded metasurface, as provided in an embodiment of the present invention. Detailed Implementation
[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0021] Currently, with the rapid development of artificial intelligence and image recognition technologies, traditional deep learning models based on electronic computing face challenges in terms of computing power and energy consumption. Electronic computing systems based on the von Neumann architecture require frequent data transfer during large-scale matrix operations, limiting processing speed and energy efficiency, and exhibiting high latency and high power consumption. Optical computing, with its high parallelism and ultra-low latency properties of photons, has become an important direction for overcoming computing power bottlenecks.
[0022] Among them, optical diffraction neural networks achieve all-optical inference by directly modulating the phase of the light field through a multi-layer diffraction structure, enabling tasks such as image classification and target detection without photoelectric conversion, thus providing a new paradigm for high-speed, low-power intelligent task processing. However, single-layer diffraction units have limited control over the light field, and existing solutions often employ multi-layer diffraction unit stacking structures, resulting in problems such as large system size and poor optical path stability.
[0023] Based on this, the image recognition simulation model and system provided in this invention can be applied to the interdisciplinary field of optical computing and artificial intelligence. Specifically, it provides an image recognition optical simulation model and system based on an optical diffraction neural network. An optical diffraction neural network architecture based on a three-layer cascaded metasurface is designed. The metasurface, as a subwavelength artificial structure array, can achieve high-degree-of-freedom control of the amplitude, phase, and polarization of the light field in a two-dimensional plane through nanostructure design. Simultaneously, the network depth is extended through multi-layer cascading to enhance feature abstraction capabilities. Multi-layer cascaded diffraction is used to achieve nonlinear mapping, directly outputting the classification light intensity distribution, which characterizes the classification result.
[0024] To facilitate understanding of this embodiment, a detailed description of an image recognition simulation model disclosed in this embodiment of the invention will be provided first.
[0025] Example 1: This invention provides an image recognition simulation model, which includes: an optical diffraction neural network for handwritten digit recognition tasks, see [link to relevant documentation]. Figure 1The diagram shows a schematic of an optical diffraction neural network, which includes an input layer, a diffraction layer, and an output layer, which are connected sequentially. The input layer modulates the incident light into a wavefront array carrying image information. The diffraction layer performs multiple diffractions on the light output from the input layer to simulate a convolution weight matrix, adjusting the phase distribution through backpropagation of errors. The output layer generates a focused light spot corresponding to the number of categories. The focused light spot focuses on the numbers corresponding to multiple detection points, and the intensity distribution of the focused light spot on the detection surface is used to characterize the classification probability.
[0026] The image recognition simulation model in this embodiment may include an optical diffraction neural network architecture based on a three-layer cascaded metasurface for handwritten digit recognition tasks. This architecture adopts a three-layer cascaded metasurface structure, including an input layer, a diffraction layer, and an output layer. The input layer modulates the incident light into a wavefront array carrying image information (e.g., MNIST images, a set of standard image datasets for handwritten digit recognition). The diffraction layer is used to realize the optical equivalence of the convolution weight matrix, perform the core computational task, and, combined with the error backpropagation algorithm, find the optimal phase distribution so that the light intensity distribution generated at the output end of the entire optical system can accurately classify the image (e.g., MNIST images). The output layer generates a focused light spot corresponding to the number of categories, focusing 10 detection points corresponding to the digits 0-9. The intensity distribution of the light spot on the detection surface directly represents the classification probability.
[0027] In some embodiments, the optical diffraction neural network described above is used to project information carried by the input image onto a preset target detection unit through training, thereby suppressing the light intensity projected onto non-target detection units.
[0028] In this embodiment, when training the optical diffraction neural network based on image recognition, the network is driven to continuously learn and project the information carried by the input image onto the correct target detection unit in a highly concentrated manner, while suppressing the sum of the light intensity of all other non-target detector units to the lowest possible level, thereby directly realizing a clear and high-contrast category discrimination signal in the target detector area.
[0029] This invention provides an image recognition simulation model. The input layer modulates the incident light into a wavefront array carrying image information; the diffraction layer diffracts the light output from the input layer multiple times to simulate the convolution weight matrix, adjusting the phase distribution through error backpropagation; the output layer generates a focused light spot corresponding to the number of categories, focusing the light spot on the numbers corresponding to multiple detection points, and the intensity distribution of the focused light spot on the detection surface is used to characterize the classification probability. In this method, high-degree-of-freedom control of the amplitude, phase, and polarization of the light field can be achieved in a two-dimensional plane through nanostructure design; at the same time, the network depth is extended by multi-layer cascading to improve the feature abstraction capability, and nonlinear mapping is achieved by multi-layer cascaded diffraction, directly outputting the classification light intensity distribution, which characterizes the classification result.
[0030] Example 2: This invention provides another image recognition simulation model, implemented based on the foregoing embodiments, focusing on the subwavelength nanostructures of an optical diffraction neural network. In some embodiments, the metasurface of the optical diffraction neural network includes multiple subwavelength nanostructures, each subwavelength nanostructure serving as a diffraction neuron for training phase values.
[0031] In this system, the output of each layer of diffractive neurons serves as the input to the next layer of diffractive neurons. Multiple diffractive neurons modulate the light field and corresponding field distribution of the input metasurface from the optical diffractive neural network.
[0032] In this embodiment, an optical diffraction neural network based on a three-layer cascaded metasurface can be set up. Each metasurface is composed of hundreds of thousands of subwavelength nanostructures, and each structural unit is a "diffraction neuron". That is, each layer contains hundreds of thousands of diffraction neurons with trainable phase values. The diffraction neurons on each layer act as point sources of secondary waves connected to the neurons of the next layer. These neurons together modulate the light field from the input surface and the corresponding field distribution.
[0033] In this embodiment, during forward propagation, when the light wave is incident on the first... When the layer is in the range, according to the Rayleigh-Sowerfeld diffraction equation, the first... Layer location nodes On the next floor The secondary light field distribution generated at node +1 for: ; in, The wavelength of the incident wave is... express With point The distance between them.
[0034] In some embodiments, the amplitude and phase of the secondary light wave output by each layer of diffractive neurons are determined based on the incident light field of that layer of diffractive neurons and the transmission coefficient of that layer of diffractive neurons.
[0035] In this embodiment, the artificial neuron node acts as a secondary light source, and the amplitude and phase of the secondary light wave generated are incident from the upper layer onto the artificial neuron node. The incident light field and the transmission coefficient of the neuron are jointly determined. Therefore, the first... The position in the layer is point On the next floor The output at the position is All points in layer -1 for points The sum of the outputs multiplied by The transmittance coefficient of the point is then multiplied by .
[0036] in, For point In the The next layer The secondary light field distribution at the location, For point In the The next layer Output at position for point In the The next layer Transmission coefficient at the location.
[0037] In some embodiments, the illumination source of the optical diffraction neural network is a complex value, which carries the phase and amplitude information of the input image. The light emitted by the illumination source is diffracted by free space light propagation and serves as the first layer input wave of the optical diffraction neural network.
[0038] See also Figure 2 The diagram shows a schematic of the implementation process of an optical diffraction neural network based on a three-layer cascaded metasurface. The illumination source of the optical diffraction neural network is a complex value, carrying the phase and amplitude information of the input image. After diffraction, the illumination source becomes the first layer input wave of the optical diffraction neural network.
[0039] In some embodiments, the optical diffraction neural network includes: multiple diffraction layers; the input wave of the target diffraction layer is determined based on the light field distribution of the target diffraction layer, the transmission coefficient of the target diffraction layer, and the output wave of the target diffraction layer; the output wave of the target diffraction layer is determined based on the input waves of multiple diffraction layers preceding the target diffraction layer.
[0040] For the diffraction layer, the first input wave of the layer Output wave and transmission coefficient As shown in the following formula, For the first The first layer 1 node For a node in the next layer of the network. These are amplitude modulation parameters. The phase modulation parameters can be calculated using the following formula: .
[0041] like Figure 2 As shown, the entire diffraction neural network can consist of 3 layers. The light field intensity that the detector on the output surface can receive is... for: .
[0042] In some embodiments, an optical diffraction neural network is used to determine a cost function based on the mean square error between the intensity of the focused spot output by the output layer and the target.
[0043] like Figure 2 As shown, the output field of view receives light field information modulated by multi-layer diffraction. The network is trained using stochastic gradient descent combined with backpropagation. To evaluate the algorithm's performance during iteration, a cost function is defined using the mean square error between the network's output intensity and the target intensity. : ;in, Let be the number of all sampling points on the output surface. The loss function is defined as the target light field. and the actual output light field The root mean square error between diffraction clusters.
[0044] In this embodiment, the phase value can be adjusted through error backpropagation during the training phase. During the iterative process of error backpropagation, a fixed amount of training data is injected into the diffraction neural network, the gradient of each layer is calculated, and the optical diffraction neural network is updated accordingly.
[0045] Example 3: This invention provides another image recognition simulation model, implemented based on the foregoing embodiments, focusing on the specific implementation methods of the image recognition simulation model and the optical diffraction neural network. See also... Figure 3The diagram illustrates the structure of an all-optical diffraction neural network. Transmission between neural network layers is achieved through light diffraction; each point on the diffraction layer is a secondary spherical wavelet source. The input to a neuron in the next layer is defined as the superposition of the outputs of all neurons in the previous layer after diffraction propagation. The weight of each neuron is defined as the phase and amplitude of the unit optical structure. Training data is input from the input layer, and the output of the neural network is calculated using optical diffraction. The phase and amplitude of each neuron are continuously optimized through backpropagation of errors. By analogy with the forward and backward propagation process of traditional neural networks, a well-trained all-optical diffraction neural network can achieve specific functions with high precision.
[0046] See Figure 4 The diagram shows a training schematic of an image recognition algorithm based on a three-layer cascaded metasurface optical diffraction neural network. The optical diffraction neural network model based on the three-layer cascaded metasurface in this embodiment is designed to automatically recognize handwritten Arabic numerals through the optical diffraction module.
[0047] In some embodiments, the image recognition simulation model further includes: a sensor array and a fully connected layer; the sensor array is used to receive the optical signal of the focused spot output by the output layer and convert the optical signal into a numerical signal; the fully connected layer is used to determine the probability of the numerical signal belonging to each category through an activation function.
[0048] In this embodiment, the sensor array can have 10 specific regions. The light intensity in each region is converted into a numerical signal, followed by a fully connected layer. The softmax function (used to convert a set of arbitrary real numbers output by a neural network into a probability distribution) is used as the activation function. The softmax function maps the outputs of multiple neurons to the interval (0, 1), which can be seen as the probability that the current output belongs to each category, thereby performing multi-class generation. ; in, This is the output value of the softmax function. For the first The output value of the class This refers to the number of categories. The MNIST dataset for handwritten digit recognition has 10 categories. The optical diffraction neural network model based on a three-layer cascaded metasurface classifies the digits 0-9. A value of 10 can be used. Record the detection results obtained after applying the softmax activation function. The largest value among them is the final test result.
[0049] Training requires the results of the annotation. The comparison and labeling results are also represented as vectors of 10 numbers. For the numbers... The annotation results for the handwritten image should only be the first one. The first bit is 1, and the remaining bits are all 0. and Error terms obtained from comparison : .
[0050] The formula for cross-entropy loss is as follows: ;in, This represents the cross-entropy loss value.
[0051] The error between the predicted value and the true label is calculated using the cross-entropy loss function. The gradient is calculated layer by layer using the backpropagation algorithm, and the phase parameters are updated using gradient descent (e.g., using the Adam (Adaptive Moment Estimation) optimizer). After several iterations on the MNIST dataset, the trained optical diffraction neural network is obtained. In the testing phase, the trained network parameters are fixed, and the test image is input into the network to perform calculations, outputting the final detection results.
[0052] Example 4: Corresponding to the above embodiments, this embodiment of the invention provides an image recognition simulation system, which is implemented based on the foregoing embodiments. The image recognition simulation system includes: the image recognition simulation model provided in the foregoing embodiments.
[0053] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process of the image recognition simulation system described above can be referred to the corresponding process in the aforementioned embodiment of the image recognition simulation model, and will not be repeated here.
[0054] Furthermore, in the description of the embodiments of the present invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in the present invention based on the specific circumstances.
[0055] In the description of this invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the referred element must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.
[0056] Finally, it should be noted that the above-described embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit it. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the technical scope disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.
Claims
1. An image recognition simulation model, characterized by, The image recognition simulation model includes: an optical diffraction neural network for handwritten digit recognition tasks, the optical diffraction neural network including: an input layer, a diffraction layer and an output layer, the input layer, the diffraction layer and the output layer being connected in sequence; The input layer is used to modulate the incident light into a wavefront array carrying image information; The diffraction layer is used to diffract the light output from the input layer multiple times to simulate the convolution weight matrix and adjust the phase distribution through error backpropagation. The output layer is used to generate a focused light spot corresponding to the number of categories. The focused light spot focuses on the numbers corresponding to multiple detection points, and the intensity distribution of the light spot on the detection surface is used to characterize the classification probability.
2. The image recognition simulation model of claim 1, wherein, The optical diffraction neural network is used to project the information carried by the input image onto a preset target detection unit through training, and to suppress the light intensity projected onto non-target detection units.
3. The image recognition simulation model of claim 1, wherein, The metasurface of the optical diffraction neural network includes multiple subwavelength nanostructures, each of which serves as a diffraction neuron for training phase values.
4. The image recognition simulation model of claim 3, wherein, The output of each diffractive neuron in the layer is the input of the diffractive neuron in the next layer. The light field and the corresponding field distribution from the input metasurface of the optical diffractive neural network are modulated by multiple diffractive neurons.
5. The image recognition simulation model of claim 4, wherein, The amplitude and phase of the secondary light wave output by each diffractive neuron in each layer are determined based on the incident light field of the diffractive neuron in that layer and the transmission coefficient of the diffractive neuron in that layer.
6. The image recognition simulation model of claim 1, wherein, The illumination source of the optical diffraction neural network is a complex value, which carries the phase and amplitude information of the input image. The light emitted by the illumination source is diffracted by free space light propagation and serves as the first layer input wave of the optical diffraction neural network.
7. The image recognition simulation model according to claim 1, characterized in that, The optical diffraction neural network includes: multiple diffraction layers; The input wave of the target diffraction layer is determined based on the optical field distribution of the target diffraction layer, the transmission coefficient of the target diffraction layer, and the output wave of the target diffraction layer; The output wave of the target diffraction layer is determined based on the input waves of the multiple diffraction layers preceding the target diffraction layer.
8. The image recognition simulation model according to claim 1, characterized in that, The optical diffraction neural network is used to determine the cost function based on the mean square error between the intensity of the focused spot output by the output layer and the target.
9. The image recognition simulation model according to any one of claims 1-8, characterized in that, The image recognition simulation model also includes: a sensor array and a fully connected layer; The sensor array is used to receive the optical signal of the focused spot output by the output layer and convert the optical signal into a numerical signal. The fully connected layer is used to determine the probability that the numerical signal belongs to each category through an activation function.
10. An image recognition simulation system, characterized in that, The image recognition simulation system includes: the image recognition simulation model according to any one of claims 1-9.